2.1. Covariation

Several types of models were considered to explore relationships: linear, quadratic, exponential and logarithmic. The model with the highest \(R^{2}\) is presented on each plot.

⚠️ Only linear models were implemented for now, as there are some bugs with the calculation of the others.

2.2. Correlation

Correlations have been calculated with Spearman’s rank coefficient.

3.1. Taxa identity

The most abundant taxa in our study area are:

• Density: B.neotena (1969), E. integra (1158), P.grandimana (1092), Nematoda (1044) and M. calcarea (575)
• Biomass: E. parma (biomass of 531.5), Strongylocentrotus sp. (65.3), N. incisa (58.5), M. calcarea (45.4) and S. groenlandicus (34.3)

The following graphs present the distribution of sampled phyla along index of cumulative exposure, according to density (left panel) or biomass (right).

Exposure categories are based on the exposure index: the higher the index, the lower the status. Maximum cumulative exposure is 1.384. Five exposure categories from ‘bad’ to ‘high’ have been set with 20 %, 40 %, 60 % or 80 % of the maximum exposure.

3.2. Community characteristics

The following graphs present the distribution of community characteristics along index of cumulative exposure.

Exposure categories are based on the exposure index: the higher the index, the lower the status. Maximum cumulative exposure is 1.384. Five exposure categories from ‘bad’ to ‘high’ have been set with 20 %, 40 %, 60 % or 80 % of the maximum exposure.

4.1. Data manipulation

For the following analyses, independant variables are exposure indices, dependant variables are community characteristics. Variables have been standardized by mean and standard-deviation.

We added abiotic parameters to the independant variables, but this did not increas significantly predictive power.

All stations and predictors were selected for the regressions, as we are interested in each of them (following graphs are for information only).

4.2. Univariate regressions

We used linear models for the regressions on community characteristics. Variables have been standardized by mean and standard-deviation (coefficients need to be back-transformed to be used in predictive models). Variable selection was not needed here, as we are interested in all exposure indices.

Results of regressions (coefficients with a significant p-value for marginal tests) are shown on the table below:

Details of the regressions, with diagnostics and cross-validation, are summarized below.

## FULL MODEL

## Adjusted R2 is: 0.21

## RMSE from cross-validation: 0.9148124

## FULL MODEL

## Adjusted R2 is: 0.02

## RMSE from cross-validation: 1.018865

## FULL MODEL

## Adjusted R2 is: -0.01

## RMSE from cross-validation: 1.014689

## FULL MODEL

## Adjusted R2 is: 0.3

## RMSE from cross-validation: 0.9270929

## FULL MODEL

## Adjusted R2 is: 0.15

## RMSE from cross-validation: 1.095543

## FULL MODEL

## McFadden's pseudo-R2 is: 0.14

## Unbiased RMSE from cross-validation: 35.999

## FULL MODEL

## McFadden's pseudo-R2 is: 0.12

## Unbiased RMSE from cross-validation: 92.58253

## FULL MODEL

## McFadden's pseudo-R2 is: 0.16

## Unbiased RMSE from cross-validation: 18.9517

4.3. Multivariate regression

The model selected by the DistLM procedure has a \(R^{2}\) of 0.22. Colours represent the value of the cumulative exposure index (the bluer, the higher).